Vehicle dynamics is sometimes portrayed as a black art. A shadowy world of empirical tuning that’s more craft than science.
The reality, of course, is quite the opposite. As far back as the 1920s, pioneers like Maurice Olley – then at Rolls-Royce – were taking a numerical approach to ride and handling. Over the years, this has given rise to increasingly sophisticated simulation techniques. Now, artificial intelligence (AI) has the potential to take data-driven engineering to new heights, delivering answers to complex questions far quicker than traditional simulation methods can.
Much of what we term AI in engineering relates to machine learning. This is a subset of AI that allows computers to learn without being explicitly programmed.
Crucially, a purely data-driven model makes no attempt to simulate the fundamental physics taking place. Neither does it use a straightforward statistical technique like averaging between two values. Instead, technologies such as artificial neural networks learn by example – analysing training data taken from physical testing or conventional simulation methods to spot the hidden relationships that exist within.
This uncanny ability can make such methods far faster than traditional simulation in some cases, and also means it can handle complex problems that are simply too difficult to represent accurately with current physics models.
“We tend to focus on what we call intractable physics – problems that are impossible or very time-consuming to solve with traditional simulations,” explains John Pasquarette, vice president for product marketing at Monolith, an AI specialist.
“There are some areas that traditional methods can simulate really well, but you end up in a Catch-22 situation. You want to increase the accuracy, so you keep tuning the simulation, but that makes it more and more complex, so you end up taking hours or even days to run a test,” he adds. “That’s where machine learning comes into play. It’s a simpler way to model the behaviour of complex systems.”
Complex problems
The number of variables in the simulation is generally a good indication of this complexity. For instance, it is unlikely that you would use machine learning to model something straightforward like the force versus extension on a spring. But when it comes to modelling climate patterns or spotting abnormal growths on a medical CT scan, the technology comes into its own. The same applies to complex problems in the automotive world, such as tyre behaviour.
Robin Tuluie, founder and co-CEO of PhysicsX, explains: “If we think about tyre performance models like Pacejka models, they’re underpinned with physics, but they have a lot of empiricism piled on top. It’s the same with brush-type viscoelastic and thermal tyre models, again where you’ve got some loose underpinning of physics and chemistry, but then ultimately a lot of empirical tuning. Data-driven models with some structural underpinnings are a natural fit for those sorts of problems.”
PhysicsX is a start-up based in Oxfordshire, UK, which is making waves in the world of engineering AI, but Tuluie himself is perhaps better known as the vehicle dynamics mastermind who led Renault and Fernando Alonso to a string of back-to-back Formula 1 World Championships in the early 2000s before repeating the feat at Mercedes with Lewis Hamilton.
He points out that it’s something of a myth that machine-learning models can simply churn their way through vast amounts of unsorted data and pick out the bits they need. Nonetheless, they do lend themselves to combining different types of data. For instance, traditional physics models for the more straightforward aspects can be integrated alongside experimental observations.
“There’s a sliding scale of models, from fully data-driven models down to physics-informed models. The latter are physics models where you’re trying to get as much as possible of the behaviour represented by structural equations and then just using loss functions to try and minimise the error. For tyre models, where the physics and chemistry are so complex, I think that sliding scale is shifting further towards the data-driven side,” notes Tuluie.
Another example he points to is the use of machine-learning models to replace aero maps in vehicle simulations, which typically take the form of look-up tables.
“By virtue of putting this data into a look-up table you’re grossly oversimplifying the CFD simulations or the wind data used to create the map. With machine learning we can now do away with the aero map approach completely, and incorporate much more detailed wind tunnel and CFD data,” he continues. “These models are dramatically quicker than CFD. They take about a tenth of a second to run, so we’re approaching the point where they could be used in real-time in a driving simulator.”
Quality not quantity
The volume of data required to train an AI model may seem off-putting, but it needn’t be huge. Somewhere in the region of 100 CFD tests, for instance, might be sufficient to piece together a data-driven aerodynamics model – or potentially a lot less if it’s supplementing a physics-based model. But the quality of this data is just as important as the quantity.
“There’s definitely a process to cleaning the data,” comments Pasquarette from Monolith. “Sometimes you just need to transform it. For instance, maybe you’ve got velocity and you just need to put a derivative on it to get acceleration. It’s not rocket science, but there can be a certain amount of manipulation needed to get the data ready to go into a model.”
Fortunately, automotive OEMs already have high-quality simulation models and well-validated test facilities, so the foundation is generally there for reliable training data. However, training the models can be quite computationally expensive. Typically, cloud computing platforms from the likes of Amazon Web Services (AWS), Oracle and Google are called in to help churn through the data. Once that’s done, though, the finished model can generally be run on a standard laptop.
Another common theme when you speak to AI experts is that the software is evolving into more user-friendly forms that don’t require the user to have a computer science background when navigating the platform.
“Almost all automotive OEMs are actively looking into integrating AI into their design workflows,” comments Srikanth Adya, lead application engineer at Ansys, a CAE/multiphysics engineering simulation software company.
“As the AI response starts getting faster, AI models will be integrated much tighter into the design tools, and they could be used to get real-time predictions on morphed designs. This leads to democratising virtual evaluations of designs by putting them in the hands of design engineers [rather than simulation specialists],” adds Adya.
Anomaly detection
As with any form of simulation, there are a multitude of different ways to harness these predictive capabilities.
“One of the applications where AI can help is in anomaly detection,” notes Pasquarette. “If you’re collecting a lot of data on a test track or a K&C rig, you don’t want to look back weeks later and realise that one of the sensors was incorrectly calibrated. We can compare that to the model as the data is collected, not just looking at the sensor readings, but also the relationships between them, to flag up anything that’s changing unexpectedly.”
AI can also help to optimise test plans by pinpointing the most important aspects to evaluate, either by extrapolating outwards or identifying gaps in the current dataset, as Pasquarette explains: “You can use these tools to step beyond your current data set and say ‘if I were to test these three conditions next, which would teach me the most, beyond the area that I’ve already covered?’ We found that by applying this approach, something like 50 to 70% of your original test plan can become redundant.”
New opportunities
There have been pockets of machine learning in the automotive industry for at least 20 years, but only now is it starting to become a recognised tool in digital prototyping. Engineers who have actively used AI for vehicle dynamics remain in the minority, but an increasing number are starting to ponder how it might be applied in this field.
David Pook is one of those keeping a close eye on this technology. An industry veteran of more than 20 years, he held a string of senior vehicle dynamics positions at Jaguar Land Rover before founding his own dynamics consultancy business, VeDynamics, in 2019.
While he has yet to use AI in an engineering capacity, Pook says there are a number of potential applications where he can see it coming into play. One in particular is attribute definition.
“When you’re developing a vehicle, you start off with subjective targets for how you would like that car to drive, from which you derive a suite of objective metrics like K&C behaviour, weight distribution and centre of gravity height. Finally, those cascade into component specifications for things like dampers, bushings and suspension links. This can be a very lengthy process,” he explains.
“One area where I think AI would be useful is if we could set out the vehicle-level behaviour that we want, and then ask AI to define the component that could deliver that, without us having to cascade through all those different levels,” adds Pook.
It’s the sheer complexity and the number of interacting variables that makes this task another good candidate for AI.
“Today, even the best objective definitions still don’t fully define the subjective driving experience,” comments Pook. “It’s all too easy to end up with objective metrics that look the same, but with vehicles that drive differently. Similarly, you can get changes at a component level that don’t produce the differences you expect at a K&C or full vehicle level, so there’s clearly something happening that you’re not measuring. There are so many metrics that the human brain can become a bit overwhelmed.”
It is possible that AI could help to build more comprehensive models that will capture more detail on the subtle variations that cause one vehicle to be perceived differently to another. But perhaps the most important factor to understand is the human brain. Here again, AI might be able to help.
“The reason we build big, expensive driving simulators is frankly because we’re not yet good enough at modelling the human driver in the loop,” points out Tuluie from PhysicsX. “Humans are incredibly sensitive and complex controllers.”
Part of the problem is that drivers are not just closed-loop controllers. As well as that ability to respond to incoming signals like changes in lateral acceleration or steering torque, experienced drivers anticipate what’s going to happen based on their past experience.
“There are some approaches with reinforcement learning that are showing good results at capturing this behaviour,” comments Tuluie. “The problem is that you require vast amounts of data to train these models.”
This blend of open- and closed-loop behaviour is also key to understanding vehicle attributes, Pook explains: “As you drive a car for the first time, your brain is building up a map of how the car is going to respond. The first few times you turn the wheel, for instance, you apply a certain steering angle at a certain speed, and from that your brain starts to predict what the car will do in other situations.”
Generally, the aim is to develop a car that will behave consistently across the whole ‘map’, but that is not as simple as it sounds.
“You might get a car that subjectively appears to behave in a very proportional fashion, but when you actually measure its steering response there’s not a straight line anywhere,” comments Pook.
“Likewise, you can end up with two cars with exactly the same yaw and lateral acceleration response, but if the steering is lighter on one car, people will always tend to perceive that as more responsive,” he adds. “Other aspects, however, are more nuanced than that, and it is possible that AI could help us to better understand that link between the objective and subjective performance of the car.”
All our experts agree that subjective evaluation and hands-on tuning will remain a major part of vehicle dynamics development for the foreseeable future. But AI could help to streamline that process, allowing engineers to focus on the job of developing better cars rather than grappling with simulation software or test plan logistics. After all, that’s where the real magic happens.
